Context Language Models let a model edit its own context
Researchers from Meta, MIT and the University of Washington introduced Context Language Models, which manage and edit their own context rather than relying on fixed summarisation, compression and retrieval machinery. The paper sets out zero-shot, in-context learning and reinforcement learning variants, each reporting accuracy gains alongside lower compute, with open questions on dropped information and the safety risk of a model curating what it remembers.
Terence Tao sets out a "Math 2.0" view of AI in mathematics
Tao published his Caltech slides arguing for a restructured mathematical practice around formalisation and machine assistance, following the week's dispute over how to assess the large batch of AI-generated results.
Microsoft's Copilot EVP sets out the product gap behind AI agents
In a Q&A, Microsoft's EVP of Copilot, Jacob Andreou, said the company rebuilt Copilot around three capabilities, Home, Code and Autopilot, after cutting back entry points that drove traffic but little real usage. He argued no single lab's model now wins at everything, so Copilot's Auto mode routes each task to whichever model suits it best while still letting people pick a specific one, and he split agent actions into irreversible "one-way doors" that always need a human in the loop against reversible ones that can run independently inside pre-computed boundaries.
Google's AMIE chatbot interviews patients before urgent-care visits
Google tested its AMIE medical chatbot with 98 real patients ahead of urgent-care appointments, letting it collect their history by text so doctors had a summary in advance. The doctor's eventual diagnosis appeared among AMIE's suggested possibilities 90% of the time, with a supervising physician monitoring every conversation and stepping in during five of them.
Manus raises more than $500m in new funding
The AI agent startup Manus closed over $500m in fresh funding, continuing a run of large rounds for general-purpose agent companies.
Norway proposes a temporary ban on AI glasses in parks, beaches and schools
Norway's government proposed temporarily banning AI-enabled smart glasses from selected public spaces, citing concerns about hidden recording.
Shield AI's autonomous X-BAT fighter draws $400m in commitments
Shield AI said total commitments for its AI-piloted X-BAT fighter aircraft have reached $400m, with flight testing expected by the end of the year.
Google's Playground turns a text prompt into a playable browser game
Google released Playground, a tool that generates a small browsable game directly from a text prompt.
ChatGPT narrows shopping down to a handful of AI-picked products
ChatGPT now lines up a short list of products with prices and reviews once a shopper states what they want and their budget, drawing on listings that stores, including Shopify merchants, send it directly. A survey of more than 1,500 US shoppers found half already turning to AI to research purchases ahead of the holidays, about three in four saying they find its picks at least somewhat trustworthy, and 16% saying they would let AI complete the purchase itself.
Odyssey's world model lets people walk through a prompt-generated place
Odyssey built a world model that turns a text prompt into an explorable first- or third-person environment that keeps predicting what happens next as the viewer moves, rather than producing a fixed video clip. Developers are already using it to train robot arms and self-driving cars inside generated worlds.
Karpathy maps the gap between light and heavy AI use as a funnel
Andrej Karpathy argued perceptions of AI vary so widely because exposure to it varies hugely: he estimates about 75% of people have barely touched it, around 20% use free chatbots occasionally, roughly 0.2% use top coding and maths tools heavily, and only around 5,000 people have access to labs' in-house systems.
McDonald's sued over claims its AI pricing tool helps franchisees coordinate prices
McDonald's faces a proposed class action alleging its AI pricing-recommendation system helps franchisees coordinate prices with each other; the company called the suit inaccurate and said franchisees set their own prices. Walmart faces a similar dispute over a 2023 patent describing shelf labels that could adjust prices based on a shopper's cart, which it says does not show the capability is in use. A Gallup survey found only 27% of Americans trust businesses to use AI pricing responsibly, down from 31% a year earlier.